are "untraditional" architechtures a bad idea?
- Thread starter Josiah
- Start date
PaulS said:
I guess I was saying that I think they have--though possibly not in the precise direction people might think when describing "non-traditional." PS2.0 support is kind of non-traditional, in the sense that the vpu becomes limited by the power of its shader engines before local bus bandwidth and other "traditional" factors weigh into the performance picture. I think that's pretty non-traditional.
I think that what happens when people take "non-traditional" approaches is that they do so based on assumptions of benefits they will reap as the result of doing so. What often happens is that those assumptions turn out to be faulty and the estimated benefit never materializes. If we go back to GigaPixel and its TBR architectures--the company long thought to be M$'s pick for supplying the graphics engine for xBox1--the chief benefit to their architecture was economic. Ram was a good deal more expensive at the time and the theory was that their TBR architecture would allow them to use much cheaper, slower ram while remaining competitive with the existing traditional pipeline renderers of the day which were far more dependent on local bus bandwidth. It wasn't so much that it was believed performance would be better, but that it wouldn't suffer while it could be obtained a good deal cheaper than traditional renderers--that's the way it seemed to me at the time. But economies of scale have a nasty habit of obsoleting economic assumptions, often before they can even be finalized.
So I think the question IHVs ask themselves is "What is the purpose for considering a non-traditional approach?" With pixel shaders, that's self evident. The challenge then is to develop a stong ps engine. So I don't think IHVs are opposed to the non-traditional at all, just that they need sound reasons to approach architectures from a non-traditional viewpoint. I'm not so sure that it's an attempt to be "safe" as it is to be "smart."
WaltC said:
That's an interesting point, but it's not only limited to non-traditional designs. The likes of nVidia and ATi have to work a couple of years ahead, so their whole business is based around assumptions. They have to assume that they're going to need X feature running at Y speed. Take the NV30 for example - If DX9 had been as slow to market (as in, to games) as previous versions, NV3x would have been fine, because they're pretty competitive in DX8 games. The assumption that DX8 would basically be "it" for the lifetime of the NV3x architecture cost nVidia dearly, whereas ATi correctly realised it's importance.
So, whilst non-traditional approaches are always riskier, i don't think the mindset going into them is any different.
PaulS said:
Well, in this case it's kind of difficult to imagine how a chip best suited for DX8.x might be considered a forward-looking, non-traditional design, wouldn't you say?
It isn't as if the IHVs are on an island as to what they estimate is needed for the future. There's a whole passle of developers from which they can divine these things, not to mention M$, and even the ARB "committe"...In very real ways the relationship between all of these parties is symbiotic--they rely on feedback and cooperation among themselves to set the pace for the future. But even so, they are individual companies and will make individual decisions for which they will be rewarded or suffer consequence. In that sense, the manufacturing and sale of any product is in some measure dependent on how well various companies read the tea leaves. Also, let's not forget the dynamic that keeps the hardware engines turning--competition. The success of a company often relates to how well it has perceived the likely directions of its competitors in terms of chip architecture, product design, and marketing. In the case of nV3x vs. R3x0, it's not that nVidia misread the progress of DX9 nearly so much as it is how nVidia misread the capability of a competitor--ATi, in this case. I think if anything ATi overestimated what nVidia was likely to do with nV30 and nVidia understimated ATi. And thus we have the present disparity. Of course, all of this presumes the engineering abilities of the two companies are relatively equal. Basically, if you want to remain the leader in an industry you have to set the pace for the future. This is something nVidia failed to do with nV3x, IMO. That's where an estimation of raw ability might enter the picture.
Heh...
I hope I haven't wandered too far afield, here...K.I.L.E.R said:
Of course!
Maybe have I misunderstood you, but you wrote:
This implied to me that you think that all "traditional" architectures are poorly executed and we thus have to move away from them. That didn't make any sense to me - just think of the R3x0/RV3x0 series: they are "traditional" architectures, but I don't regard them as poorly executed.
I was talking about all the current non-traditional architectures in the paragraph.
I have yet to see a non-traditional architecture that is/was successful as a traditional one.
IE: The Geforce architecture (meaning every product based on the original Geforce 256 architecture) was far more successful than the Kyro architecture.
traditional = single chip, immediate render mode and no TBR.
non-traditional = multi chip, defered renderer with TBR and any other unique architectural features.
Joe DeFuria
Legend
Ailuros said:
True...
But they are often intimately tied. It's a lot tougher to market and distribute a product that was poorly executed.
Ailuros said:
Marketing plays the largest role in the execution of the product.
Don't forget reliability. I have yet seen any reliable non conventional architecture in the desktop segment of the market.
Don't forget that the companies who launch these "non-traditional" products have a strict budget, I bet if given a good budget that Ati or nVIDIA could release a non conventional architecture that will whipe the floor with conventional architectures currently in the market.
Voodoo 5 Dual chip = late, low yeilds, low budget.
Kyro/2 = low budget = low end part
Ati Rage Fury MAXX (sp?) = Ati's past was plagued by unreliability, poor marketing, poor performance and poor expectations.
I don't believe the architectures themselves are to blame. A Kyro 2 with a T&L unit, couple of VS and some solid good performing drivers would be quite a steal.
It's usually sales/marketing that get products out the door and make products look good/bad.
It's the accounting dept that are responsible for finance.
I believe that those 2 areas are the main problems.
PR and budget.
Have I gone off track again?
No not really. And I can't say that I disagree either with your points in general. Only other difference is a matter of semantics or definitions if you prefer.
The first thing that pops into my mind when I read bad execution are things like buggy hardware or a bad transition from design to final hardware (whichever sounds better).
The highlighted part is what I meant. On the K2 paradigm itself, I heard years ago that the original STG5000 (4*1, ~166MHz, DDR, T&L) was cancelled because the price/performance ratio wasn't as good as compared to a higher clocked KYRO (always open to corrections if things were different). Now there's a chance that this was either ST Micro's decision or a joint decision between ImgTec and ST; in any case there never should have been a K2 initially at least IMHO. The tradeoff would have come later on when the first full T&L games arrived and it would have given the card a much longer shelf-life.
Plus Series3 was plagued by some delays too.
Anyway did I say that I agree to that perspective?
So what you guys think of Sony/IBM CELL thing?
A link to a patent.
http://appft1.uspto.gov/netacgi/nph...)&OS=an/"sony+computer"&RS=AN/"sony+computer"
How is it more cool than PC parts?
A link to a patent.
http://appft1.uspto.gov/netacgi/nph...)&OS=an/"sony+computer"&RS=AN/"sony+computer"
How is it more cool than PC parts?
Yea in the end, all that matters is price/performance. Which is why there is still hope for interesting competition in the (IMO) rather dull Gfx scene these days.
Nv3x and ATI are both traditionals in the old sense of the word. Both are doing well from a market standpoint. Nv3x looks like it suffered from feature creep, a very powerful chip with loads of features and bandwidth, yet suffering from lack of balance in a few puzzling areas. i'm going to wait for nv4x before dismissing Nvidia though, the amount of talent they have on staff is still impressive. Still ATI has the better overal part at this time.
Everyone on this forum has been waiting patiently for ImgTech to decide to hit the high end, or for another Gigapixel to start up and finance themselves with a nice bank loan. Almost everyone knows the theoretical advantages on paper that tilers have, and its frustrating no company has at least experimented with trying to take on the big 2, who look locked with traditional designs.
One huge problem of course are drivers. A new tiler will have to be both backward compatible with old dx versions, as well as having sufficient driver depth to expose their features to a decidedly pro traditional API in DX. As time goes on this becomes progressively harder and harder to do. Not to mention the days of bandwidth is king (one of the tilers strong points), are becomming fuzzier (witness Nv3x for instance).
What happened to BitBoys was also depressing, I for one would have liked to have seen embedded memory solutions try to compete.
Nv3x and ATI are both traditionals in the old sense of the word. Both are doing well from a market standpoint. Nv3x looks like it suffered from feature creep, a very powerful chip with loads of features and bandwidth, yet suffering from lack of balance in a few puzzling areas. i'm going to wait for nv4x before dismissing Nvidia though, the amount of talent they have on staff is still impressive. Still ATI has the better overal part at this time.
Everyone on this forum has been waiting patiently for ImgTech to decide to hit the high end, or for another Gigapixel to start up and finance themselves with a nice bank loan. Almost everyone knows the theoretical advantages on paper that tilers have, and its frustrating no company has at least experimented with trying to take on the big 2, who look locked with traditional designs.
One huge problem of course are drivers. A new tiler will have to be both backward compatible with old dx versions, as well as having sufficient driver depth to expose their features to a decidedly pro traditional API in DX. As time goes on this becomes progressively harder and harder to do. Not to mention the days of bandwidth is king (one of the tilers strong points), are becomming fuzzier (witness Nv3x for instance).
What happened to BitBoys was also depressing, I for one would have liked to have seen embedded memory solutions try to compete.
But supposedly, another strong point for a tiler (or maybe i should say TBDR) is that it "doesn't need to shade what you don't see". So not only will it have a bandwidth advantage, it'll also have a shader advantage. Something that is rather important these days. (witness Nv3x for instance
)
Bjorn said:
But supposedly, another strong point for a tiler (or maybe i should say TBDR) is that it "doesn't need to shade what you don't see". So not only will it have a bandwidth advantage, it'll also have a shader advantage. Something that is rather important these days. (witness Nv3x for instance
Are IMRs rendering everything concerning Pixel shaders? I can see an advantage up to now for TBDRs (theoretically at least) what the fact concerns that PS and VS are perfectly de-coupled. Other than that without carefully addressing the Vertex Shader side on a TBDR it can actually turn into a disadvantage. I've no idea of course what can be done for the latter.
Must be the reason why it's sooooo fast with FP32? Ever thought of that?
Bjorn said:But supposedly, another strong point for a tiler (or maybe i should say TBDR) is that it "doesn't need to shade what you don't see". So not only will it have a bandwidth advantage, it'll also have a shader advantage. Something that is rather important these days. (witness Nv3x for instance
But there's this idea that a TBR product doesn't have to do much of anything to do that...
It's computationally dependent for the gpu to only render the pixels that can be seen. This requires a gpu overhead just like ps engines do. I think one of the main reasons TBR's never took off is that it was discovered that it was cheaper and faster to do brute-force rendering with z-occlusion techniques than to accomplish TBR. TBR doesn't eliminate overdraw in software--it doesn't rely on bandwidth like BFRs to handle overdraw--but it still has to handle it in the gpu itself, which is done through its tiling techniques. Up to now, at least in the performance sector, it's been the case that TBR computational overhead exceeds its bandwidth benefits, and BFRs have therefore remained dominant. IE, it hasn't mattered that BFRs render many more pixels than needed than TBRs, because the BFR can still render frames faster. IMO, TBRs made much more sense when ram was ~$50 a meg than they do now.Up til now specific implementations have proven they couldnt compete ... all the rest from any of us is speculation. At least I assume you didnt develop simulators and mathematical models to test your assumptions anymore than I did for mine WaltC
Clockspeed is a design issue ... given that the latter PVRs were coupled to DRAM when DDR was dirtcheap, with their clock synced to the low speed of the RAM, and their transistor counts rather low I dont think they were exactly meant to compete with the products in the performance sector which existed when they finally came out.
Clockspeed is a design issue ... given that the latter PVRs were coupled to DRAM when DDR was dirtcheap, with their clock synced to the low speed of the RAM, and their transistor counts rather low I dont think they were exactly meant to compete with the products in the performance sector which existed when they finally came out.
MfA said:Up til now specific implementations have proven they couldnt compete ... all the rest from any of us is speculation. At least I assume you didnt develop simulators and mathematical models to test your assumptions anymore than I did for mine WaltC
Clockspeed is a design issue ... given that the latter PVRs were coupled to DRAM when DDR was dirtcheap, with their clock synced to the low speed of the RAM, and their transistor counts rather low I dont think they were exactly meant to compete with the products in the performance sector which existed when it finally came out.
Well, my only assumption is that TBR is far less dependent on local bus bandwidth than is BFR, because of the difference in how overdraw is dealt with. And I only draw my conclusions about its performance from implementations that have been brought to market, as opposed to speculations. I think it's significant that although lots of companies have looked at TBR, few have brought such a product to market, and of those who have none has been able to compete in performance with traditional designs.
I have always liked the theory of TBR--a lot. It's just that in the years since GigaPixel I've seen no stellar implementations in comparison with traditional designs, and have reached the conclusion that they are probably unlikely (though it would be really cool to see otherwise.)
A lot of ideas spring up of economic necessity which are later discarded because economies of scale render them moot--like the concept of AGP texturing, for instance--which was a grand idea when Intel formulated the spec because videoram was ~$75 a meg and it was difficult to conceive of 16mb 3D cards, let alone the monsters so common today, and so texturing out of system ram was considered expedient and practical. But the concept of AGP texturing was always an economic one, not an ideal. I see similar considerations for TBR as being important when the concept was originally developed. Because of economies of scale the bandwidth issues relative to overdraw are of much less concern today for traditional renderers than at the time GigaPixel was working on its chips, for instance.
Similar threads
